benchr
Something is not working
What goes wrong when you actually use this, why it happens, and what to do instead. Every record says what no fix solves.
Unless a record is marked otherwise, a vendor page or a published paper explains the cause.
For
Nothing matches that. Clear the filters to see everything again.
Documents and files
Output you can use
It stopped doing what I asked
- Halfway through, it stopped following the instructions I gave at the startRe-state the two or three rules that matter immediately before the request, not at the top of the conversation. Recency beats position.
- It gets most of the way through, then just stops mid-sentenceAsk for one section, not the whole thing, and then ask for the next. Two short answers beat one truncated one.
- Identical prompt, different provider, completely different behaviorBefore blaming the prompt, check that the feature it depends on exists on the target model at all. Half of these failures are a missing parameter, not a wording problem.
Images
- The poster looks great until you read the words on itKeep the text to a few words, put it in straight quotation marks in the prompt, and say where on the image it goes.
- Same character, second image, different faceStop re-describing the person and start re-supplying them: pass the first image back in as a reference for every subsequent generation.
- The eyes change color and the kid looks older on every pageName the attributes that must not move — eye color, hairline, one distinctive mark — and check those three specifically rather than judging the face as a whole.
Video
- The product looks different in every shot of my adGenerate from a still of the actual product rather than from a description of it, for every single shot.
- The video shimmers — textures crawl and the light flickersShorten the shot. Most shimmer becomes invisible below about two seconds, and a cut costs nothing.
Agents and automation
- My coding agent fixed the bug and broke three other thingsWork on a branch, and read the diff before the tests. A large diff for a small request is the signal, not the test result.
- It scores brilliantly on the demo and falls over on my actual workTake ten tasks from your own week, write down what a correct answer looks like, and run those. That is your benchmark; everyone else's measures someone else's work.
- It described the chart correctly and then got the number wrongGive it the numbers, not a picture of the numbers. A CSV export, the underlying table, or the data endpoint removes the entire failure mode.
- The browser agent clicks the wrong buttonRefer to targets by their accessible name, not by where they are. 'The button labeled Continue' survives a layout change; 'the button on the right' does not.
Cost
Limits and interruptions
If you are seeing an API error
Seeing an actual API error code? Those are catalogued separately.
See all